基于对抗训练的快速图像实例分割研究

Research on Fast Image Instance Segmentation Based on Adversarial Training

面向复杂交通路况图像的快速目标检测和实例分割技术是实现无人驾驶的关键,现有两阶段分割方法通过多步骤执行检测分割任务,耗时长且在目标定位后存在大量特征处理。为此,提出一种结合生成对抗网络和目标检测的快速实例分割模型(FISAT)。首先在目标检测网络中引入并行分支生成掩码,对每个交通实例对象进行分割;然后加入ROI类损失对每类掩码进行学习并用感知损失保存掩码图像信息;最后使用谱归一化解决生成对抗网络训练过程中的慢收敛问题。在MSCOCO基准测试上,FISAT目标分割每秒帧(FPS)可达到47.0,为MNC和FCIS的5倍。在分割优化上,使用Darknet提取器的FPS达到43,相较于Resnet提取器提高8.0。在平均精度(AP)上,FISAT相较于两阶段Mask-RCNN提高7%,相较于一阶段方法平均提高24%。

Fast target detection and instance segmentation technology for complex traffic images is the key to achieve autonomous driving. The existing two-stage method performs the detection and segmentation task in multiple steps, which takes a lot of time and is difficult to optimize.The one-stage method has a lot of post-processing after target positioning, which is difficult to meet the real-time requirements. To solve the above problems, propose a fast instance segmentation model(FISAT) based on target detection and generative adversarial network. Firstly, the parallel branch generation mask is introduced into the target detection network to segment each traffic instance object. Secondly, ROI class loss is added to learn each class mask, and perceptual loss is used to save mask image information. Finally, spectral normalization is applied to solve the slow convergence problem in the training process of generative adversarial network. On the benchmark of MSCOCO, the frame per second(FPS) of FISAT can achieve 47.0, which is five times that of MNC and FCIS. In terms of segmentation optimization, the Darknet extractor FPS achieves 43, which is 8.0 higher than that of the Resnet extractor. The average accuracy(AP) of FISAT is 7% higher than that of the two-stage Mask-RCNN and 24% higher than that of the one-stage method.